A machine learning-based SMS/Email spam classifier built with Streamlit and deployed on Heroku. The project uses Natural Language Processing (NLP) techniques and Naive Bayes classification to accurately identify spam messages.

Project Overview

This mini project demonstrates end-to-end machine learning workflow from data exploration to model deployment:

• Dataset: SMS Spam Collection with 5,572 messages
• Accuracy: 97.20%
• Algorithm: Multinomial Naive Bayes with TF-IDF vectorization

Dataset

• Source: SMS Spam Collection Dataset
• Total Messages: 5,572 (5,169 after removing 403 duplicates)
• Distribution:
  • Ham (Legitimate): 87.37% (4,516 messages)
  • Spam: 12.63% (653 messages)
• File: spam.csv

Class Imbalance

The dataset is imbalanced with significantly more ham messages than spam. The model handles this through appropriate algorithm selection (Naive Bayes performs well on imbalanced text data).

Project Structure

SpamDetection_MiniProject/
├── app.py                          
├── sms-spam-detection.ipynb        
├── model.pkl                       
├── vectorizer.pkl                  
├── spam.csv                       
├── requirements.txt                
├── Procfile                          
├── setup.sh                        
├── nltk.txt                        
└── .gitignore                   

Installation & Setup

Prerequisites

• Python 3.7+
• pip or conda package manager

Local Installation

1. Clone the repository

   git clone <repository-url>
   cd SpamDetection_MiniProject

2. Install dependencies

   pip install -r requirements.txt

3. Download NLTK data (required for text preprocessing)

   python -c "import nltk; nltk.download('punkt'); nltk.download('stopwords')"

4. Run the application

   streamlit run app.py

5. Access the app Open your browser and navigate to http://localhost:8501

Usage

1. Launch the Streamlit app
2. Enter your SMS/Email message in the text area
3. Click the "Predict" button
4. The app will classify the message as either Spam or Not Spam

Example Messages

Spam Example:

WINNER!! You have been selected to receive a $1000 prize.
Call now at 555-0123 to claim your reward!

Ham Example:

Hey, are we still meeting for lunch at 1pm today? Let me know!

Model Development Process

1. Data Cleaning

• Removed 403 duplicate messages
• Dropped unnecessary columns (Unnamed: 2, 3, 4)
• Encoded target labels (ham=0, spam=1)

2. Exploratory Data Analysis (EDA)

• Character Analysis: Spam messages are longer (avg 137 chars) vs ham (avg 70 chars)
• Word Count: Spam has more words (avg 27) vs ham (avg 17)
• Sentence Count: Spam has more sentences (avg 3) vs ham (avg 2)
• Word Clouds: Visualized most common words in spam vs ham messages

3. Text Preprocessing Pipeline

1. Lowercase conversion
2. Tokenization (word_tokenize)
3. Remove special characters (keep alphanumeric only)
4. Remove stop words & punctuation
5. Porter Stemming (reduce words to root form)

Example Transformation:

Input:  "I'm gonna be home soon and i don't want to talk about this stuff anymore tonight"
Output: "gon na home soon want talk stuff anymor tonight"

4. Feature Engineering

• TF-IDF Vectorization: max_features=3000
• Converts text into numerical vectors representing term importance
• Alternative features tested: character count, word count, sentence count

5. Model Selection & Evaluation

Models Tested

Algorithm	Accuracy	Precision	Notes
Multinomial Naive Bayes	97.20%	100%	Best - Selected for deployment
Extra Trees Classifier	97.68%	99.15%	High performance but more complex
Random Forest	97.49%	98.28%	Excellent but slower
SVC (Sigmoid Kernel)	97.29%	97.41%	Good but slower training
XGBoost	97.00%	94.21%	Good but overkill for this problem
AdaBoost	97.20%	95.04%	Comparable to MNB
Logistic Regression	96.13%	97.12%	Simple and effective
Gradient Boosting	94.87%	92.93%	Slower with marginal gains
Bagging Classifier	96.81%	86.15%	Lower precision
Decision Tree	94.39%	83.81%	Overfitting risk
K-Nearest Neighbors	92.84%	77.12%	Poor performance on text

Why Multinomial Naive Bayes?

• 100% Precision: No false positives (crucial - don't want to block legitimate messages)
• 97.20% Accuracy: Excellent overall performance
• Fast Training & Inference: Real-time predictions
• Low Memory Footprint: Small model size (model.pkl is only 96KB)
• Proven for Text Classification: Industry standard for spam detection

Model Performance Metrics

Confusion Matrix (Test Set)

                Predicted
              Ham    Spam
Actual Ham    896      0     ← Perfect! No legitimate messages marked as spam
     Spam      29    109    ← 79% spam detection rate

Technical Stack

Machine Learning & Data Science

• scikit-learn: Model training, evaluation, and vectorization
• pandas: Data manipulation and analysis
• numpy: Numerical computations
• nltk: Natural language processing (tokenization, stemming, stopwords)
• XGBoost: Gradient boosting framework (tested)

Visualization

• matplotlib: Plotting and visualizations
• seaborn: Statistical data visualization
• wordcloud: Word cloud generation for EDA

Web Application

• Streamlit: Interactive web interface for model deployment

Deployment

• Heroku: Cloud platform for application hosting
• pickle: Model serialization

Dependencies

streamlit
nltk
scikit-learn

Additional packages used in notebook:

• pandas
• numpy
• matplotlib
• seaborn
• wordcloud
• xgboost

Learning Outcomes

This project demonstrates:

1. End-to-End ML Pipeline: From raw data to deployed application
2. Text Classification: NLP techniques for real-world problem
3. Model Selection: Comparing 10+ algorithms and choosing optimal solution
4. Handling Imbalanced Data: Strategies for skewed class distributions
5. Feature Engineering: TF-IDF vectorization for text data
6. Web Deployment: Creating interactive ML applications with Streamlit
7. Cloud Deployment: Hosting on Heroku with proper configuration